Phonemic segmentation using the generalised Gamma distribution and small sample Bayesian information criterion

In this work, we present a text-independent automatic phone segmentation algorithm based on the Bayesian Information Criterion. Speech segmentation at a phone level imposes high resolution requirements in the short-time analysis of the audio signal; otherwise the limited information available in such a small scale would be too restrictive for an efficient characterisation of the signal. In order to alleviate this problem and detect the phone boundaries accurately, we employ an information criterion corrected for small samples while modelling speech samples with the generalised Gamma distribution, which offers a more efficient parametric characterisation of speech in the frequency domain than the Gaussian distribution. Using a computationally inexpensive maximum likelihood approach for parameter estimation, we evaluate the efficiency of the proposed algorithm in M2VTS and NTIMIT data sets and we demonstrate that the proposed adjustments yield significant performance improvement in noisy environments.